Evidence locker · first written 18 September 2026 · updated 9 October 2026 · Denver
How the search is governed: what the machine is allowed to do on its own, where it has to stop, how the two AI models split the work between them, and what all of it writes down while nobody is watching.
The machine researches, scores and drafts. It never decides which opportunities to pursue, and it never sends anything.
The machine steps on the left run unattended, except Gemini’s review, which Ben carries across by hand. Every crossing to the right is a person, and nothing moves back left until that person has ruled.
The machine is two AI models with separate jobs. Claude builds and measures. Gemini reviews what Claude built, as the model that did not build it, and owns how it reads. Neither settles a disagreement on its own. Ben decides.
Claude
Builds, measures and keeps the record
Neverrewrites Gemini’s wording, and never counts its own check as the review.
Ben
Decides, and reviews last
Neverdelegated: a decision, a send, or the last look.
Gemini
Reviews the work, and owns how it reads
Neverrestates or certifies a number, a name, a codename or a date. Its review is a judgment against the brief; facts are checked at their source.
Every asset a third party will see, every new claim about Ben’s history, and every change to how the search scores, guards or tiers a role makes this trip before it ships or takes effect. The brief it is checked against is written down: the playbook, the search parameters and the master resume.
01 · Claude
Builds the draft and measures it
Figures read from their sources, counts computed, the file read back from where it lives.
02 · Claude
Writes one self-contained handoff
The facts Gemini may use, the hard limits, the part of the brief it is checked against, numbered slots, the shape of the return. The text travels inside it, never a link.
03 · Ben
Carries it into a fresh Gemini chat
No saved instructions, so nothing from an earlier project steers the answer.
04 · Gemini
Reviews it against the brief
As the model that did not build it. Returns three parts, kept separate: the text alone; every change and why; its challenges, each naming the line.
05 · Claude
Files the return word for word
The day it lands, with a written disposition on every challenge. None is dropped silently.
06 · Ben
Rules, and reads last
Any disagreement arrives as one choice in two columns, the strongest case for each side. His review is the final sweep.
Ben arbitrates
Neither model resolves a disagreement with the other. It goes to Ben as a choice, each side argued at its strongest.
Read the file, not the label
A claim about what a document says counts only if it came from reading that document, not a title, a summary or the other model’s account of it.
One source of truth
The shared Drive folder is the record, and the newest dated file wins. Gemini reviews from the handoff, not the folder, so what it judged is exactly what it was sent.
Nobody writes as Ben
Either model may draft in his register and must say so. Neither invents experience, metrics, logos or dates.
Exceptions are rulings too
The lanes bend only when Ben rules it, and the ruling says how far.
A review is not a verification
A pass from Gemini means the work meets the brief in a careful reader’s judgment. Whether a figure is true is settled at its source, and only there.
Drawn from how this search has actually been run since 27 August, not from how it ought to be run.
The division of labour
The machine researches, scores and drafts. It never decides which opportunities to pursue and never sends anything. Every send is gated on a human ruling.
Rulings, not preferences
A decision is recorded with a date and with what it superseded. The newest ruling wins. Work is never silently corrected.
The error ledger is a control, not an apology
Every failure is logged with the rule it earned, and the entry names who caught it. 144 entries so far — one hundred and eight caught by the controls, thirty-six by the person running it.
What gets automated
Anything determinable, verifiable or buildable without a judgment call. Questions are reserved for the calls only the owner can make.
Quality control runs on the reporting too
A status note that contradicts the current plan is treated as a defect, not a wording problem.
Filed in the same action that produced it
Every artefact lands where the team can find it as it is made. Nothing is handed back to a person to file.
Cadence
The active priority is worked to completion. Secondary material is reviewed on a fixed slot rather than whenever it surfaces.
The most load-bearing rule in the system, because it is what makes a score arguable rather than asserted.
Ratings, status, next action, and rulings.The judgment calls, and only the judgment calls.
Rows, and machine-verifiable fields.Anything it can check against a source it can name.
Score, Confidence, Tier.All three are computed from the rubric. Change a rating or change a weight — never the result.
A high score is not a recommendation. Four rules sit between the score and the tier, and they exist to stop the system rewarding a row it simply knows too little about.
Fewer than four criteria rated caps the row at the third tier, however high the score reads.
A seniority rating of two or below forces the third tier or lower. A manager-level title caps at the fourth.
A compensation rating of five, taken from a band the employer published, waives the thin-row floor. Published evidence beats missing evidence.
A gate, not a score. Remote or Denver passes. Unknown caps the tier until somebody resolves it.